Ph.D. Research
My Ph.D. research focuses on the evaluation of mappings of parallel application
tasks onto the nodes of heterogeneous (federated) clusters. The main goal
is to facilitate the search for efficient mappings both initially and during
the evolution of a computation. This is done essentially by providing a
comparison function which allows the comparison of mappings on the basis
of execution time cost/benefit to an application. The novelty of the present
approach is that the the comparison takes into account static and dynamic
information from both the computing system and the application and favors
the mapping that is a better match of the application's computation and
communication patterns with the available system resources.
The culmination of this research is the Cost/Benefit Estimating Service
(CBES). The prototype utilizes LAM/MPI,
XMPI, and NWS. The following is a schematic of CBES operations:
CBES can be seen as the synthesis of three components: the System Profiler,
the Application Profiler, and the Core Module.
-
The System Profiler (SP) has the duty of maintaining a current view of
the availability of main resources (CPU and interconnecting network bandwidth)
for the cluster. The SP does this by periodically measuring characteristic
load parameters on each one of the N nodes of a cluster (with the help
of a modified version of NWS) and expressing the availability of resources
at any given instant t as a deviation from the maximum, zero-load
state availability. The deviation is estimated with the help of an empirical
model of the effect of load on the availability, with input the periodically
measured load parameters and output the deviation from the zero-load state.
Thanks to this model the O(N²) problem of knowing at moment t
the available bandwidth between any pair of cluster nodes becomes an O(N)
problem.
-
The Application Profiler (AP) creates a database with detailed computation
and communication profiles for applications of interest. Each such profile
is obtained by executing an application under specific conditions and storing
a trace of its execution. Subsequently, this trace gets analyzed with a
modified version of the XMPI tool, which can provide detailed statistics
about the computation and communication pattern of each process (task)
of the application in a machine readable format.
-
The CBES Core Module (CM) combines the information provided by the SP and
the AP and given a number of different mappings for the application tasks
the CM generates a score for each mapping. These scores practically express
the efficiency of the corresponding mappings: the smaller the score the
faster the application executes.
CBES publications
Dimitrios Katramatos, Steve J. Chapin.
A Scalable Method for Predicting Network Performance in Heterogeneous Clusters.
Proceedings of the 2005 IEEE International Symposium on Parallel Architectures,
Algorithms, and Networks (ISPAN 2005), pp. 288-295, Las Vegas, Nevada,USA, December 7-9, 2005.
[pdf]
Dimitrios Katramatos, Steve J. Chapin.
A Cost/Benefit Estimating Service for Mapping Parallel Applications on Heterogeneous Clusters.
Proceedings of the 2005 IEEE International Conference on Cluster Computing (Cluster 2005),
Boston, Massachusetts, September 26-30, 2005.
[pdf]
Dimitrios Katramatos, Marty Humphrey, Cheol-Min Hwang, and Steve J. Chapin.
Developing a Cost/Benefit Estimating Service for Dynamic Resource Sharing
in Heterogeneous Clusters: Experience with SNL Clusters.
Proceedings of the 1st IEEE/ACM International Symposium on Cluster
Computing and the Grid (CCGrid 2001), pp. 355-362, Brisbane, Australia,
May 15-18, 2001.
[Postscript]
Dimitrios Katramatos, Deepak Saxena, Nehal Mehta, and Steve J. Chapin.
A Cost/Benefit Model for Dynamic Resource Sharing.
Proceedings of the 9th Heterogeneous Computing Workshop (HCW 2000),
Cancun, Mexico, May 1-5, 2000.
[Postscript]